用于高效思维链推理的 Logit-熵自适应停止启发式方法
Logit-Entropy Adaptive Stopping Heuristic for Efficient Chain-of-Thought Reasoning
AI总结:
本文提出无需训练的 LEASH 解码算法,通过监控 token 级熵斜率和 top-logit 边际改善两个信号自适应停止思维链生成,在 GSM8K 和 AQuA-RAT 上以约 10 个百分点准确率下降为代价,将 token 生成量减少 30--35%、延迟降低 27%。
摘要:
思维链(CoT)提示是使大语言模型能够进行复杂推理的关键技术。然而,生成完整、固定长度的推理过程在计算上是浪费的,既增加了 token 使用量,也提高了延迟。我们提出了 LEASH:Logit-熵自适应停止启发式方法,这是一种无需训练的解码算法,能够自适应地停止推理过程的生成。LEASH 监控两个内在信号:token 级熵的斜率以及 top-logit 边际的改善情况。一旦这两个信号都趋于平稳,表明模型已达到稳定的推理状态,LEASH 就会终止生成。在 GSM8K 和 AQuA-RAT 基准上对四个指令微调模型进行的实验中,LEASH 将平均 token 生成量减少了 30--35%,延迟降低了 27%,同时相对于 CoT 产生了 10 个百分点的准确率下降。LEASH 与模型无关,不需要额外的训练或监督,为 CoT 解码提供了一种简单而高效的替代方案。
英文摘要:
Chain-of-Thought (CoT) prompting is a key technique for enabling complex reasoning in large language models. However, generating full, fixed-length rationales is computationally wasteful, inflating both token usage and latency. We introduce LEASH: Logit-Entropy Adaptive Stopping Heuristic, a training-free decoding algorithm that adaptively halts rationale generation. LEASH monitors two intrinsic signals: the slope of token-level entropy and the improvement in the top-logit margin. It terminates the generation once both signals plateau, indicating the model has reached a stable reasoning state. Across four instruction-tuned models on the GSM8K and AQuA-RAT benchmarks, LEASH reduces average token generation by 30--35% and latency by 27%, while incurring a 10 p.p. accuracy drop relative to CoT. LEASH is model-agnostic and requires no additional training or supervision, offering a simple and efficient alternative to CoT decoding.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。